Deepmind's AlphaProof Nexus solves 9 Erdős problems for $100s

google deepmind introduced alphaproof nexus, an agentic framework combining gemini 3.1 pro with the lean formal verifier to autonomously solve mathematical conjectures. the system successfully resolved nine erdős problems and 44 oeis conjectures, achieving results at an inference cost of just a few hundred dollars per problem. the architecture relies on a loop where the llm proposes proof steps and the lean compiler checks them, grounding the model in symbolic logic to avoid reasoning errors. surprising researchers, the simplest agent configuration using only compiler feedback proved sufficient for the hardest tasks, signaling a shift toward generalist llm-based approaches over specialized systems.